Alphabetical orthographies

Pollatsek et al. (1992) were first to show evidence for phonological PB during reading of English sentences. They found that parafoveal previews that are homophonous with the target word (e.g. “bare” as the preview of “bear”) led to significantly shorter first fixation durations compared to an orthographic control condition (e.g., “bean” as the preview of “bear”). This effect was 20 ms in size and was taken as evidence that the phonological information from the parafovea is preserved across fixations in order to aid foveal word recognition. Further support for this conclusion was found by Miellet and Sparrow (2004), who reported that French pseudo-homophone previews (e.g., “cheise” as the preview of “chaise”) also resulted in shorter fixation durations compared to visually similar pseudo-word control previews (e.g., “choise” as the preview of “chaise”). While explicit statistical comparisons between pseudo-homophone and pseudo-control previews were not reported, the condition means were consistent with the pattern of results obtained by Pollatsek et al. (1992). Because pseudo-homophones do not exist in participants’ phonological lexicon, Miellet and Sparrow (2004) interpreted their findings as evidence that phonological codes are assembled parafoveally before the target word is fixated.

Chace et al. (2005) further investigated how reading skill may modulate the effect. They found that if they excluded participants who scored lower than the 20-percentile rank on the Nelson-Denny Reading Test, there was evidence of a significant phonological PB effect in gaze durations. The authors interpreted this as a replication of Pollatsek et al. (1992) and motivated the exclusion of low-scoring participants by arguing that “…such a high concentration of low-level readers is not indicative of the general student population (Chace et al., 2005, p. 211)”. Additionally, when only data from the homophone and orthographic control conditions were considered, “more skilled” readers (≥60 percentile) showed a robust phonological PB effect in gaze durations (a mean difference of 35 ms), whereas “less-skilled” readers (≤40 percentile) showed no evidence for phonological PB (a difference of –15 ms). Chace et al.’s results suggest that only skilled adult readers benefit from phonological information in the parafovea. However, another interpretation may be that poor readers add noise to the data, which makes it difficult to detect small effects. Because no theory to our knowledge would predict a negative difference instead of a zero difference in the less-skilled group, this could indicate that the study had non-negligible measurement error.

Bélanger, Mayberry, and Rayner (2013) investigated the effect in skilled and less-skill deaf readers, as well as in a control group of hearing adults. While neither the skilled nor the less-skilled deaf readers showed evidence for phonological PB, the hearing controls did show evidence for the effect in first fixation duration (24 ms). However, this effect was only apparent when the phonological preview word was high frequency and the target word was low frequency. Bélanger et al. (2013) also considered this to be a replication of previous findings (Chace et al., 2005; Miellet & Sparrow, 2004; Pollatsek et al., 1992), arguing that the frequency of parafoveal words may influence the activation of letter identities, and therefore the processing of phonological information in the parafovea. Interestingly, however, most target/homophone preview pairs used in their study were borrowed from Pollatsek et al. (1992), who reported a general phonological PB effect irrespective of lexical frequency. Another experiment by Choi and Gordon (2014), which also partly borrowed from Pollatsek et al.’s (1992) targets, found no evidence for phonological PB with either homophone or pseudo-homophone previews.

Tiffin-Richards and Schroeder (2015) investigated the parafoveal processing of phonology in child and adult German readers by comparing pseudo-homophones (e.g., Bläch- Blech) to orthographic controls (e.g., Blüch- Blech). While children showed phonological PB effects in single fixation duration (28 ms) and gaze duration (34 ms), the adults did not. Tiffin-Richards and Schroeder (2015) argued that the lack of an effect in adults may be due to the greater orthographic transparency of German and the fact that their reading stimuli were optimized for children. More recently, Blythe, Dickins, Kennedy, and Liversedge (2018) tested the parafoveal processing of phonology in English teenagers with permanent hearing loss, as well as in reading- and age-level hearing controls. Interestingly, all three groups showed a significant phonological PB effect with peseudo-homophone previews in gaze durations, thus suggesting that they were able to activate the phonological codes of the parafoveal word.

To examine whether the phonological PB can still be observed even in the absense of orthographic similarity, Jouravlev and Jared (2018) tested English-Russian bilinguals. They presented Russian preview words written in the Cyrilic alphabet that had phonological but little or no othographic overlap with the English target words (e.g., “бланк [blank]– blood”) and compared them to previews that had neither phonological nor orthographic overlap (e.g., “гжель [ɡʐɛlʲ]–blood”). Jouravlev and Jared (2018) reported a significant phonological PB effect independent of orthography for both first fixation duration (20 ms) and gaze duration (43 ms). More recently, Leinenger (2018) also examined the phonological PB in English readers. She found significant evidence for phonological PB in English with pseudo-homophone previews (Experiment 4), but not with homophones previews (Experiment 3). In both experiments, the size of the effect was comparable, although it was considerably more modest than that of previous studies (6–8 ms).

To summarise, classical studies have shown that readers of alphabetical orthographies can process phonological information parafoveally (Miellet & Sparrow, 2004; Pollatsek et al., 1992). However, a closer look at the literature shows that this effect may be less consistent than what is sometimes believed. While some recent studies have replicated the effect (Blythe et al., 2018; Leinenger, 2018, Experiment 4), others have either not replicated it (Choi & Gordon, 2014; Leinenger, 2018, Experiment 3) or have found it only after data exclusions based on factors, such as the reading skill of participants (Chace et al., 2005) or the lexical frequency of the preview and the target word (Bélanger et al., 2013). While not explicitly stated, some of the modulation analyses could have been done post-hoc, as they were not always incorporated in the study design. This raises questions about the generalisability of the effect across different populations and reading situations.

It is also interesting to note that some of the early studies have found phonological PB effects that are similar in size to the classical PB effect (valid vs invalid preview difference; see Figure 1a for an illustration), which has been estimated to be 26 ms on average for first fixation duration and 37 ms for gaze durations (Vasilev & Angele, 2017). As the source of the classical PB effect is assumed to be some combination of abstract letter and phonological codes (Rayner, White, Kambe, Miller, & Liversedge, 2003), this raises the possibility that these early studies may have overestimated the size of the phonological PB effect. Estimating the exact size of the effect is important for designing experiments that are sufficiently powered to detect it.

Mandarin Chinese

Most studies on the phonological PB have been conducted in alphabetical orthographies such as English. However, as phonology is not always represented in the same way across different languages, there could be cross-linguistic differences in the effect. For example, languages with non-alphabetical orthographies such as Mandarin Chinese differ from English in a number of ways that could potentially influence the parafoveal processing of phonology (see Tsang & Chen, 2012; Yu & Reichle, 2017). Unlike English, Chinese uses a logographic writing system where characters, rather than letters, are the basic unit. Chinese is an unspaced language because individual characters are not demarcated by spaces and there are no boundaries between words. Most words in Chinese are one or two characters long, although some words can consist of four or more characters (X. Li, Zang, Liversedge, & Pollatsek, 2015). Chinese characters are square-shaped and are made up of basic strokes (Wang & Yang, 2007). Some of the strokes can be combined to form radicals and each character consists of at least one such radical.

Around 80% of all Chinese characters are compound characters, which are made up of a phonological and one or more semantic radicals (Chen, 1996; Shu, 2003). The semantic radical is usually located on the left side of the character and the phonological one is located on the right side of the character, although this is not a fixed feature of the language and their positions can vary (Chen, 1996). While the semantic radical provides information about the meaning of the character, the phonological one provides a guide to its pronunciation, though not always a very reliable one (Shu, 2003). This is because phonological information is more loosely encoded in Chinese characters due to the very long history of the language during which multiple phonologies of individual character components have emerged (Honorof & Feldman, 2004). For example, Wengang (1991) took all compound characters from a dictionary and calculated that the average reliability with which phonetic radicals represent the pronunciation of their characters is only 0.58. This is the case because only 36% of phonetical radicals correctly represent the pronunciation of compound characters, 48% partially represent the pronunciation, and 16% do not represent the pronunciation of the character at all (Wengang, 1991). Therefore, phonetic radicals do not consistently provide reliable information regarding the pronunciation of such compound characters.

Liu et al. (2002) were first to investigate the parafoveal processing of phonology during the reading of Chinese. They reported that a phonologically similar character preview (e.g., 药 as the preview of 要) led to significantly shorter gaze durations and total viewing time compared to an unrelated preview character that was neither phonologically nor orthographically similar to the target (e.g., 位 as the preview of 要). However, this result was significant only when the authors analysed fixation durations on a target area that included the target, pre-target, and post-target character. Tsai et al. (2004) replicated the phonological PB effect in gaze durations and further extended Liu et al.’s (2002) results by showing that visually similar homophones also lead to a significant phonological PB in first fixation duration. Tsai et al. (2004) also found that characters with highly consistent phonology (i.e., the ones where the phonetic radical consistently represents the same pronunciation across different characters) showed phonological PB in first fixation duration, whereas those with lower consistency did not. The authors argued that this last result suggests that Chinese readers can access the phonological information from parafoveal vision and use it in the early stages of character identification.

More recently, Yan et al. (2009) found a numerical trend that was consistent with the phonological PB effect in gaze durations with simple Chinese characters, but the effect did not reach significance. Interestingly, Tsai, Kliegl, and Yan (2012) adapted Yan et al.‘s (2009) stimuli for readers in Taiwan (where traditional Chinese characters are used), but failed to find a significant phonological PB or any indication of an effect in that direction. Nevertheless, post-hoc analyses indicated that the effect became significant with increasing preview time (i.e., longer gaze durations on the pre-target word), but only when that word was of high frequency. These early results suggest that, while the effect may exist in Mandarin Chinese, it may be weaker under certain circumstances.

Pan et al. (2016) further investigated how reading mode (silent vs oral) may modulate the effect. Because oral reading requires the articulation of words, Pan et al. predicted that the phonological PB should be larger in oral compared to silent reading due to the greater relevance of phonological information in the former mode. Contrary to this prediction, Experiment 1 found a significant phonological PB in silent, but not in oral reading, using the same simple character previews from Yan et al.‘s (2009) study. Nevertheless, as there was a significant phonological parafoveal-on-foveal effect in oral reading, the authors argued that the results are still consistent with the notion that phonological processing is greater during oral reading. Additionally, Experiment 2 provided direct evidence for this hypothesis using compound characters, as the phonological PB was significant for both first fixation duration and gaze duration in oral reading, but not in silent reading (although there was a numerical trend in second-pass measures). Therefore, these results suggest that there may be greater parafoveal processing of phonology during oral reading, at least for compound characters.

More recently, Luo, Wu, and Jiao (2018) investigated the role of phonetic radicals on the parafoveal processing of phonology. In line with the results of Tsai et al. (2004), they found that readers were able to process phonology at the radical level, which was shown by a significant difference between characters with regular phonology and characters with irregular phonology in first fixation durations. In this study, regularity referred to whether the compound character had the same pronunciation as its phonetic radical or not. Interestingly, however, no phonological PB was found at the character level, because a compound character that was homophonous with the target character but did not share the same phonetic radical did not differ from an unrelated character. This last result is contrary to some previous findings (e.g., Liu et al., 2002; Pan et al., 2016) and further highlights the need for a systematic analysis of the phonological PB effect.

In summary, despite the differences with which phonology is represented in Chinese, there is some evidence indicating that readers may be able to activate phonological information parafoveally (Liu et al., 2002; Luo et al., 2018; Pan et al., 2016; Tsai et al., 2004). Nevertheless, there are also some inconsistencies in the results. Future research may verify that the effect is moderated by additional factors, such as the type of characters used and/or the reading mode (Luo et al., 2018; Pan et al., 2016), or properties of the pre-target character and preview time (Tsai et al., 2012). However, we are interested in determining the size of the general effect (i.e., averaged over all moderating factors). As there are also a number of inconsistencies in the results from alphabetical orthographies, it is unclear whether the size of the phonological PB effect may differ between Mandarin and English.

Literature Search

The search of the literature followed the PRISMA guidelines (Moher et al., 2009) and covered all articles published until December 21, 2018. This was done by searching Google Scholar and the Web of Science with the terms “parafoveal preview AND reading” and “phonological preview AND reading” (two separate searches). In addition to this, the reference list of screened articles and their citations in Google Scholar were checked for relevant studies. The reference lists of recent literature reviews on the topic were also checked (Cutter et al., 2015; Leinenger, 2014; Schotter et al., 2012).

The study selection process is illustrated in Figure 2a. The abstracts of identified records were first screened for relevant studies and then any duplicates were removed. The full-texts were then screened for articles that did not fit the scope of the meta-analysis. Finally, the remaining articles were evaluated against the inclusion criteria presented in Appendix A. In short, the studies had to use the boundary paradigm (Rayner, 1975) to manipulate the phonological preview of a target word in skilled adult readers, have a suitable control condition and reasonable eye-tracking data quality. As a result of this evaluation, 24 studies were included in the meta-analysis (72.7% of all considered experiments). Three more eligible studies were located from the survey of the unpublished literature (see below), thus bringing the total number of included studies to 27. More information about the included studies is presented in Table 1 for alphabetical studies and in Table 2 for Chinese studies.

A Survey of the Unpublished Literature

To estimate the proportion of unpublished studies on the topic and to gather information about them, a survey was conducted that targeted researchers who have done relevant previous work. Seventy-eight researchers were identified based on their previous publication record and were sent an email invitation to participate in an anonymous survey. Every attempt was made to invite all relevant researchers who have worked on this or similar topics and to make the sample as internationally representative as possible. Participants were asked to indicate whether: 1) they have worked on this topic before and 2) have any unpublished studies. If they indicated that they have unpublished studies, participants were asked if they can provide more information- either details that will allow these studies to be included in the meta-analysis (e.g., descriptive statistics) or simply the number of unpublished studies that they have (to be used in the statistical simulations). Participants also had the option to provide no information at all. The survey was conducted between December 12, 2018 – January 31, 2019.

The results of the survey are presented visually in Figure 2b. The survey response rate was 48.7%. From the researchers who took part, 34% indicated that they have done previous research on the topic. Out of the researchers who had done relevant research, 53.8% indicated that they have some unpublished data (7 individuals). Three out of the seven researchers provided effect sizes for one unpublished study each, which made it possible to include three more experiments into the meta-analysis. To avoid duplicate entries, participants were asked if they were the lead researcher on the project, and all indicated that they were. These studies met the same inclusion criteria from the search of the published literature. From the remaining four researchers with unpublished data, one researcher indicated that they have one unpublished study but provided no effect sizes, and three researchers provided no information at all. In summary, the survey located three unpublished studies that were included in the meta-analysis and found evidence for at least four more missing studies for which no further information could be obtained. The missing studies were used in statistical simulations to evaluate the robustness of the main results.

Effect Size Calculation

Two dependent measures were used in the present meta-analysis: first fixation duration (FFD) and gaze duration (GD). FFD is the first fixation on a word during first-pass reading, while GD is the sum of all fixations on a word until readers move on to another word. The size of the phonological PB effect was calculated in milliseconds for each study. This was done by subtracting fixation durations in the phonological preview condition from the control condition (orthographic preview for alphabetical studies and unrelated word preview for Chinese studies). Because Chinese has a deep orthography and a higher number of homophones compared to languages such as English (see Duanmu, 2006), it is easier to create phonological previews that have little orthographic overlap with the target word. As a result, an orthographic control condition is not necessary in this language. This is likely why researchers have generally opted for an unrelated-word control condition that is neither phonologically nor orthographically related to the target. Positive effect sizes indicate evidence for the phonological PB, whereas negative effect sizes indicate evidence against it.

The variance of effect sizes was calculated using Formulas 12.8–12.9 from Borenstein (2009). More specifically, the variance of each study (V) was calculated as:

where S_diff is the standard deviation of the mean difference for that study and n is its sample size. S_diff was calculated as: where S₁ is the standard deviation of the phonological preview condition, S₂ is the standard deviation of the control condition, and r is the correlation between the means in the phonological and the control preview conditions. Because r could not be calculated for all studies, the population correlation between the phonological and the control condition (ρ) was estimated from six experiments for which the raw data was available. The six correlations were then meta-analysed to obtain a weighted estimate of ρ (Szalma & Hancock, 2011). The resulting estimate of ρ (0.64 for FFD and 0.62 for GD) was used for calculating the variances in all studies.

Publication Bias

In the present meta-analysis, 33.3% of all included studies were not published. To assess the data for publication bias, funnel plots were used to visually examine the effect sizes for evidence of asymmetry (Sterne et al., 2011). In the absence of bias, effect sizes are expected to form a symmetrical funnel shape when they are scattered against the inverse of their standard error, with more precise studies appearing narrowly at the top of the plot and less precise studies scattering more widely at the bottom (precision increases with the inverse of the standard error). If studies are missing from one side of the plot and the funnel shape appears asymmetric, this could indicate the presence of publication bias. The visual inspection of the funnel plots indicated no clear evidence of asymmetry (see Appendix B). To test this statistically, Egger’s regression method was used (Egger, Smith, Schneider, & Minder, 1997). The results indicated no significant evidence for funnel plot asymmetry in either FFD (t(14) = 1.70, p = 0.11) or GD (t(17) = 1.33, p = 0.19). The effect sizes from Chinese studies were not analysed statistically for funnel plot asymmetry because such analyses are not recommended when there are fewer than 10 studies in the analysis (Sterne et al., 2011).

Data Analysis

Posterior sampling

Sampling of the posterior distribution was done with JAGS v. 4.30 (M. Plummer, 2003) in the R software v.3.5.1 (R Core Team, 2018). Ten Markov Chain Monte Carlo (MCMC) chains were run with 105 000 iterations each. The starting values were randomly generated, but checks were done to ensure that this did not influence the results. The first 5000 iterations were discarded as burn-in. Therefore, the summary of the posterior distribution was based on 1 million samples. Model convergence was assessed with Gelman and Rubin’s (1992) convergence diagnostic and with visual inspection of the parameter trace-plots. The evidence suggested that all models had converged. The effective sample size, or the size of the MCMC chains after adjusting them for auto-correlation (Kass, Carlin, Gelman, Neal, & Carlin, 1998), was always at least 10 000, as recommended by Kruschke (2015) (mean: 114 951; range: 13 190–437 168).

The results are reported as the estimate of the phonological PB effect in milliseconds and its 95% credible interval. Unlike confidence intervals, credible intervals (CrI) have the intuitive interpretation that they contain the true parameter of interest with a 95% probability because the values inside this interval correspond to 95% of the posterior distribution (see Morey, Hoekstra, Rouder, Lee, & Wagenmakers, 2016). Two probability measures are also reported for the meta-analysis results. The first one is the probability that the phonological PB is bigger than 0, given the data, i.e., p(Phon. PB > 0|Data). This is taken as evidence of how likely it is that the effect exists. The second one is the probability the phonological PB is bigger than 3 ms, given the data, i.e. p(Phon. PB > 3|Data). This is a more stringent criterion that quantifies the evidence in support of a small, positive effect. The cut-off of 3 ms was chosen because a potential phonological PB effect smaller than this number may be difficult to measure reliably. Because some experiments have used eye-tracker sampling frequency of 500 Hz, effects smaller than 2 ms may be too small to reliably capture with that resolution. Therefore, we chose a cut-off that is 1 ms larger than the lowest sampling resolution (i.e., 2 + 1 = 3 ms). For more information on Bayesian methods and their interpretation, readers are referred to Nicenboim and Vasishth (2016).

Meta-analysis

The results from the meta-analysis are shown in Table 3. Additionally, the results are visualised in Figure 3 for alphabetical studies and in Figure 4 for Chinese studies. The estimated phonological PB effect measured in FFD was very small for both alphabetical (1 ms) and English-only studies (1.9 ms). This suggest that readers of alphabetical orthographies obtain little, if any, benefit from phonological previews during the first fixation of the target word. This conclusion was also supported by the finding that the probability of a true effect in FFD was only 73% for English studies. Interestingly, however, the phonological PB effect was much larger for Chinese studies in FFD (5.8 ms) and there was a 93% probability that the effect exists. Therefore, this suggests that it is very likely that Chinese readers benefit from a phonological preview during the first fixation of the target word.

The analysis of GD showed a modest phonological PB effect for both alphabetical (3.5 ms) and English-only studies (4.5 ms). The probability of a true effect was very high in both cases (≥93%), thus suggesting that it is very likely that readers of alphabetical orthographies benefit from phonological previews in GD. Interestingly, the phonological PB measured in GD was about 3 times as large for Chinese studies (14.1 ms) and the probability of a true effect was equally as high (96%). This suggests that Chinese readers are also very likely to benefit from phonological previews in GD. However, the uncertainty around this estimate was much larger. This was likely due, at least in part, to the smaller number of studies included in this analysis. However, the between-study variance (τ²) was also much larger compared to the other analyses, thus suggesting that the included studies differed between each other to a greater extent. The probability that the effects are larger than 3 ms, given the data, was more than 90% only for the Chinese effect in GD. However, the same probability for English in GD and Chinese in FFD was around 70%. This suggests that there was smaller probability that the effects are at least 3 ms in size.

A sensitivity analysis using the leave-one-out method was conducted to test the robustness of the meta-analysis results (see Greenhouse & Iyengar, 2009). In the leave-one-out method, the meta-analysis is repeated by omitting one different study each time. The results (presented in Supplementary file 1) indicated that the effect in GD for all languages and in FFD for Chinese studies was robust and not unduly influenced by any individual study. The effect in FFD for alphabetical and English studies was less robust, but the evidence for a phonological PB in this measure was not very strong to begin with.

Meta-regression

Meta-regression analyses were used to test whether the phonological PB differs between alphabetical and Chinese studies, and between homophone and pseudo-homophone previews. The meta-regression slope from this analysis indicates the estimated mean difference between the two groups of studies after adjusting it for the precision of individual experiments. There was an 88% probability that the phonological PB is larger for Chinese compared to alphabetical orthographies in FFD (mean difference: b = 2.4 ms; 95% CrI: –1.8, 6.5 ms) and a 97% probability for such difference in GD (mean difference: b = 5.3 ms, 95% CrI: –0.1, 10.7 ms). Likewise, there was an 82% probability that the phonological PB effect is larger for Chinese compared to English studies in FFD (mean difference: b = 2.1 ms, 95% CrI: –2.5, 6.47 ms) and a 95% probability that it is also larger in GD (mean difference: b = 4.8 ms, 95% CrI: –0.83, 10.3 ms). Therefore, these results suggest that the size of the phonological PB is larger in Chinese compared to alphabetical orthographies.

A separate model indicated that the phonological PB was 1.57 ms larger for homophone compared to pseudo-homophone previews in FFD (95% CrI: –3.5, 6.6 ms) and 0.66 ms smaller in GD (95% CrI: –5.4, 4.2 ms). The probability of a true advantage for homophone previews was 74% for FFD and 38% for GD. Therefore, the results suggest that there was generally no difference in the amount of benefit that readers of alphabetical orthographies obtain from homophone and pseudo-homophone previews.

Statistical Simulations of Missing Studies

While the statistical analyses indicated no evidence for the presence of publication bias, the survey of the unpublished literature nevertheless showed that there are additional unpublished studies for which no further information could be obtained. Although this does not in itself indicate the presence of bias, it does suggest that there is additional information that could not be accessed in the present meta-analysis. In light of this, statistical simulations were carried out using the survey findings to test the robustness of the meta-analysis results. Only the effect size in GD was analysed because it was generally larger.

Three different scenarios were used to simulate different types of missing data that may exist (see the top row of Figure 5 for an illustration). First, a “null effect” scenario assumed that all missing studies come from a normal distribution with a mean of 0 ms and a standard deviation of 7.5 ms. This is the most conservative scenario in which all missing studies are generated from a null effect and 95% of all studies fall between –15 and 15 ms. A second “small effect” scenario assumed that all missing studies come from a normal distribution with a mean of 5 ms and a standard deviation of 7.5 ms. In this scenario, the missing studies are generated by a small (5 ms) effect and 95% of all studies lie between –10 and 20 ms. A final “larger effect” scenario assumed that all missing studies come from a normal distribution with a mean of 10 ms and a standard deviation of 7.5 ms. In this case, the studies are generated by a larger (10 ms) effect and 95% of all studies lie between –5 and 25 ms. The variance of effect sizes was kept constant across the three scenarios; it was sampled from a normal distribution with a mean of 100 ms and a standard deviation of 200 ms that was truncated at 5 ms as its lower bound (the upper bound was +∞). The distribution was truncated at 5 rather than 0 ms because variances very close to 0 are highly unlikely given the typical precision of studies on this topic. This distribution was used because it is a good approximation of the variability of parafoveal preview effect sizes typically observed in the literature (see Vasilev & Angele, 2017).

The general method for the simulations was as follows. For each of the k missing studies, an effect size T_k and a variance of this effect size $S_k^2$M6 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ S_k^2 \] \end{document} were randomly drawn from the respective probability distributions. $S_k^2$M7 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ S_k^2 \] \end{document} was always drawn from the same distribution: $S_k^2 ~ \text{Normal}(100, 200) \text{I}[5, \infty ]$M8 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ S_k^2 \sim {\rm{Normal}}(100, 200){\rm{I}}[5,\infty] \] \end{document} . The distribution from which T_k was drawn differed based on the simulation scenario (T_k ~ Normal(0,7.5) [“null effect”]; T_k ~ Normal(5,7.5) [“small effect”]; T_k ~ Normal(10,7.5) [“larger effect”]). The k effect sizes and their variances were then added to the available data and the meta-analysis model was fit. This procedure was repeated 10 000 times for each scenario, thus yielding 10 000 simulated effect size estimates of the phonological PB. The simulations were done once with the number of known missing studies (k = 4; see Figure 2b) and once with the estimated missing studies. The estimated percentage of missing studies was calculated by dividing the number of known missing studies by the total number of researchers who have done work on this topic (4/13 = 30.7%). The number 4 was taken because one researcher indicated that they had one unpublished study and three more researchers said that they had unpublished studies, but did not indicate how many. Therefore, the number of estimated missing studies was k = 6 for alphabetical orthographies, k = 5 for English and k = 2 for Chinese. It should be noted that these simulations assume that all researchers who failed to provide information about their unpublished studies in the survey have only one relevant unpublished experiment. This represents the lowest possible bound for how many unpublished studies exist.

The results are illustrated in Figure 5 for English studies and the results for all other analyses are shown in Table 4. The phonological PB effect measured in GD was generally robust to the missing studies for all languages. The smallest effect size for alphabetical orthographies remained positive even in the most conservative “null effect” scenario. However, the average effect in the simulations was somewhat smaller in this scenario, thus suggesting that the true phonological PB effect for alphabetical studies may be even smaller if all or most of the missing studies happen to be null results. A similar result was obtained for English-only studies, but the difference was that the effect was about 1 ms larger than that of alphabetical studies. There was more variability in the simulated effect size for Chinese studies, which was likely due to the smaller number of available experiments for that language. Nevertheless, the effect size was always at least 9 ms on average. In summary, the simulations revealed that the phonological PB is robust to the missing studies, but the actual size of the effect may be slightly smaller if most of them are null results.

Statistical Power Simulations

The meta-analysis revealed that there is a high probability that the phonological PB in GD is bigger than 0 (see Table 3). Therefore, one important question is to determine the number of subjects and items needed to reliably detect the effect (see Brysbaert, 2019; Brysbaert & Stevens, 2018). To do that, statistical power simulations were used where Linear Mixed Models (LMMs) were fit using simulated, but realistic data. This was done using the “lme4” R package v. 1.1-18-1 (Bates, Machler, Bolker, & Walker, 2014). In these models, GD was the outcome variable and experimental condition (phonological vs. control preview) was the only predictor variable. Random intercepts, as well as random slopes for experimental condition were added for both participants and items (Baayen, Davidson, & Bates, 2008; Barr, Levy, Scheepers, & Tily, 2013). The results were considered statistically significant if the p-values were ≤0.05. The simulations were carried out using the “simr” R package v.1.0.5 (Green & Macleod, 2016). The variance-covariance matrix of the random effects and the residual variance were estimated using the data from Leinenger (2018), Experiment 4 for English and Yan et al. (2009) for Chinese studies. Both datasets have a reasonable number of subjects and items; therefore, it was possible to simulate realistic data that can capture the by-subject and by-item variability that is present in the effect.

In the power simulations, the size of the phonological PB effect in GD was kept constant (4.5 ms for English and 14 ms for Chinese) while the number of statistical subjects and items was being manipulated. For each subject/item number permutation, 200 simulated LMMs were run and the phonological PB effect was calculated. The statistical power of the simulation was defined as the number of LMM models that showed a significant phonological PB effect divided by the total number of models run (e.g., 100 significant models out of 200 would yield a statistical power of 0.5). If more than 5% of the models in a given simulation yielded a warning or a convergence failure, the simulation was repeated until the proportion of such models was below 5%. If this did not happen within 5 attempts, the random slopes for experimental condition were removed to help the model converge (this occurred only 0.52% of the time).

The results are presented in Figure 6. For English, the effect could be detected with 80% power with at least 90 subjects and 50 items per condition. This suggests that, while the effect very likely reflects a true difference in the population, only high-precision experiments may be able to reliably capture it. Not surprisingly, the larger effect in Chinese was easier to detect and only 40 subjects and 30 items per condition were needed to achieve 80% statistical power.

Implications for the Strong versus Weak Phonological Viewpoints

As noted above, there are two competing viewpoints on the role of phonology during visual word recognition. According to the strong phonological view (Frost, 1998), the phonological code is a mandatory component in visual word recognition. That is, phonology mediates the activation between letter/graphemes and the orthographic word form. On the other side of the debate is the weak phonological view (e.g., Rastle & Brysbaert, 2006), which posits that the phonological code is not mandatory as the orthographic code can be accessed directly from incoming letter level activation.

Masked phonological priming in visual word recognition has long been considered as compelling evidence for the strong phonological view, and the effect has been observed using several different priming paradigms. The relevance of the present research is that it provides additional evidence in favour of the strong phonological view. As the results show, there is a modest but reliable phonological PB effect. It is important to emphasize that the veracity of the strong phonological view does not depend on the size of the phonological PB effect, but rather on its existence. Indeed, given the relatively small phonological priming effects in visual word recognition (Rastle & Brysbaert, 2006), a small phonological PB effect is also to be expected, particularly when syntactic and semantic constraints on word recognition come into play during text reading. The fact that there is a phonological PB effect at all adds further support to the view that word recognition and reading require phonological activation.

Implications for Future Studies

The present results also have important implications for future studies investigating the parafoveal processing of phonology. For example, the power simulations indicated that a high number of subjects and items may be required to achieve desirable levels of statistical power, at least for alphabetical orthographies such as English. Additionally, future studies investigating how the effect may be modulated by different factors (e.g., participant population, individual differences) should consider the general size of the effect, as interaction terms may be even harder to detect. This may be particularly important when studying rare populations, such as deaf or dyslexic readers, where large sample sizes may be difficult to achieve in practice. In such cases, it may be desirable to seek converging sources of evidence- for example, from phonological priming in foveal vision where the effects may be larger.

While the present results are consistent with the existence of a small phonological PB effect, they do not tell us whether larger effects may be observed under some conditions. For example, one interesting recent hypothesis is that stronger phonological PB effects may be observed in oral compared to silent reading due to the need for articulation in the former reading mode (Pan et al., 2016). While this hypothesis has only been tested in Chinese so far, a similar effect would also be predicted in alphabetical orthographies such as English. Additionally, the phonological PB effect may be modulated by individual differences such as reading skill level. This possibility was first suggested by Chace et al. (2005), who reported that only skilled adult readers were able to process phonology parafoveally. However, the way that participants were divided into “skilled” and “less skilled” readers in their study was somewhat arbitrary and it may be beneficial to replicate their results independently. Nevertheless, as there is a growing number of studies indicating that the amount of parafoveal processing may depend on reading skill level (Veldre & Andrews, 2014, 2015a, 2015b, 2016b), this remains an interesting question to study in the future.

Because the present study had access only to the experiment-level effect sizes, it was not possible to examine whether the phonological preview benefit may be modulated by different linguistic properties of the target words, such as their lexical frequency (Risse & Kliegl, 2012, 2014; Risse & Seelig, 2019; Schotter & Leinenger, 2016; Schotter, Leinenger, & von der Malsburg, 2017; Schotter, von der Malsburg, & Leinenger, 2019) or plausibility (Schotter & Jia, 2016; Veldre & Andrews, 2016a). Therefore, this remains a task for future studies. To facilitate such analyses, we encourage researchers to routinely make their data openly available (Gilmore, Kennedy, & Adolph, 2018; Martone, Garcia-Castro, & Vandenbos, 2018; Vanpaemel, Vermorgen, Deriemaecker, & Storms, 2015).

Implications for Computational Models of Reading

Parafoveal processing is often thought to be a key part of skilled adult reading (e.g., Reichle & Reingold, 2013) and has played an important role in the development of computational models of eye-movement control, such as E-Z Reader (Reichle et al., 1998), SWIFT (Engbert et al., 2005), and more recently OB1 (Snell et al., 2018). Nevertheless, even though all three models allow for parafoveal processing, neither E-Z Reader nor SWIFT explicitly implement how different types of linguistic information, such as phonology, orthography, or semantics may be processed parafoveally. Even the more recent OB1 model (Snell et al., 2018), which includes open-letter bigrams that are activated by the visual input and subsequently excite word-level nodes, would only be capable of capturing linguistic effects to the extent that they correlate with the open-letter bigram codes. Therefore, currently no models of eye-movement control during reading explicitly explain how phonology is processed parafoveally, even though such accounts do exist in (foveal) single-word recognition (e.g., Coltheart et al., 2001; Perry et al., 2007). Rather, because the classical PB is often thought to be related to the activation of orthographic and phonological codes (Rayner et al., 2003), any model that can account for this effect is also by association consistent with the existence of the phonological PB.

Nevertheless, future models of reading may explicitly add components for the parafoveal processing of phonology, much in the same spirit as OB1 (Snell et al., 2018) has done for orthography. The present research places important constraints on future models by suggesting that only a limited amount of phonological processing may occur parafoveally, at least in alphabetical orthographies such as English. If the classical PB is indeed the result of orthographic and phonological activation (Rayner et al., 2003), then the present results suggest that the contribution which is unique to phonology is modest.

The limited size of the phonological PB effect is not necessarily surprising as the effect is not very large in single word recognition studies. For example, Rastle and Brysbaert (2006) found that phonological masked priming effects in visual word recognition are about 10–13 ms in size. Because the phonological prime in such studies is viewed foveally, the results may represent a “theoretical maximum” as to how much phonological processing can occur. In contrast, parafoveal processing necessarily incurs costs due to the decrease in visual acuity, and as a result, the effects are likely to be smaller. This is consistent with computational models of reading in which the amount of processing is modulated by the eccentricity of the parafoveal word from the point of fixation. This is perhaps more obvious in models of spatially distributed processing where either the visual input (OB1; Snell et al., 2018) or the lexical processing rate of words (SWIFT; Engbert et al., 2005) are a function of the eccentricity from the fixation point. However, even sequential-attention models such as E-Z Reader assume that the time to complete the first stage of lexical processing (L1) is modulated by the mean absolute distance between each of the letters in the word and the gaze position (Schotter, Reichle, & Rayner, 2014). Therefore, existing models are consistent with the notion that phonological processing should be smaller in parafoveal compared to foveal vision. In this sense, given the loss of visual acuity in the parafovea, the present results suggest that phonological PB effects may be smaller in practice than the effects observed in foveal phonological priming studies (see Rastle & Brysbaert, 2006).